search

TuringLang / Bijectors.jl

Implementation of normalising flows and constrained random variable transformations
https://turinglang.org/Bijectors.jl/
MIT License
200 stars 33 forks source link

Fix type instabilities of SimplexBijector #174

Closed devmotion closed 3 years ago

devmotion commented 3 years ago

Fixes type instabilities discovered in https://github.com/TuringLang/Turing.jl/issues/1276.

torfjelde commented 3 years ago

I'll review, but quick quesh first: is there a reason why you re-introduced the Val? I don't think that's necessary here; at least when I experimented, it wasn't.

devmotion commented 3 years ago

I could not obtain type stability without it. IMO this seems reasonable since it is not possible to infer the type, or rather the type parameter proj, of SimplexBijector from a boolean value proj.

torfjelde commented 3 years ago

For the case when proj is a runtime parameter you mean, or nah? It is inferrable if you do SimplexBijector{true}().

devmotion commented 3 years ago

Sure, you can hardcode it for the default version with true - but the three argument version, with true or false, is still type unstable.

devmotion commented 3 years ago

BTW also with your suggestion in https://github.com/TuringLang/Turing.jl/issues/1276#issuecomment-813074909 invlink(d, x) is still type unstable. It seems you accidentally checked link(d, x) instead?

torfjelde commented 3 years ago

Haha, yeah you're right! Sorry about that. I think I was just cycling through my command history to get the function call and seems like I was a bit too quick.

But regarding the 3-argument version. The reason why we replaced Val with proj was because when you use link and invlink in a different scope than Main, you'll have to hardcode proj to get type-stability no matter if we're using Val or not. And if you hardcode, even when not using Val, you should get type stability due to constant propogation.

Thouhgh in the light of the stuff we saw in the DPPL PR, I'm not so certain about that anymore:) And it seems like even something as simple as

f(d, x) = Bijectors.invlink(d, x, true)

fails :confused:

So I guess I'm in favour of adding back the Val.

devmotion commented 3 years ago

Yes, IMO this is a breaking change.